A simulation comparison of phylogeny algorithms under equal and unequal evolutionary rates [published erratum appears in Mol Biol Evol 1995 May;12(3):525]

Using simulated data, we compared five methods of phylogenetic treeestimation: parsimony, compatibility, maximum likelihood, Fitch-Margoliash, and neighbor joining. For each combination of substitutionrates and sequence length, 100 data sets were generated for each of 50trees, for a total of 5,000 replications per condition. Accuracy wasmeasured by two measures of the distance between the true tree and theestimate of the tree, one measure sensitive to accuracy of branch lengthsand the other not. The distance-matrix methods (Fitch- Margoliash andneighbor joining) performed best when they were constrained from estimatingnegative branch lengths; all comparisons with other methods used thisconstraint. Parsimony and compatibility had similar results, withcompatibility generally inferior; Fitch- Margoliash and neighbor joininghad similar results, with neighbor joining generally slightly inferior.Maximum likelihood was the most successful method overall, although forshort sequences Fitch- Margoliash and neighbor joining were sometimesbetter. Bias of the estimates was inferred by measuring whether theindependent estimates of a tree for different data sets were closer to thetrue tree than to each other. Parsimony and compatibility had particulardifficulty with inaccuracy and bias when substitution rates varied amongdifferent branches. When rates of evolution varied among different sites,all methods showed signs of inaccuracy and bias.